Open h-vetinari opened 2 years ago
Currently tensorflow here is not built against MKL (OneDNN), but do we plan to?
Currently tensorflow here is not built against MKL (OneDNN), but do we plan to?
Not sure how involved this is, but since onednn seems to be available for all relevant platforms, I'm not opposed to trying!
Tensorflow is building oneDNN from source in the build currently. That is probably the first 1-2h of CI. Getting that from the feedstock instead of building from source would help a lot.
Tensorflow is building oneDNN from source in the build currently. That is probably the first 1-2h of CI. Getting that from the feedstock instead of building from source would help a lot.
That sounds like a big win if we can pull it off!
Since rebuilding takes so much effort and we don't do it that often, I wanted to create an issue to track the various things we should adapt for the next build:
EDIT: "checked" below means it's currently included in #176
bazel-toolchain
(see https://github.com/conda-forge/bazel-toolchain-feedstock/pull/5)cudatoolkit
as a dependency totensorflow
, or even better, get rid oftensorflow-base
(since it's not needed anymore now thattf-estimator
is built in this feedstock), see https://github.com/conda-forge/tensorflow-feedstock/issues/162numpy
has been unpinned from 1.19.2 in TF 2.7 as of https://github.com/tensorflow/tensorflow/commit/a90383a867a2337600c775d67842f308da9d118b__cuda
virtual package or the GPU builds, see #174gilbertfrancois
from the maintainers: https://github.com/conda-forge/tensorflow-feedstock/issues/178libtensorflow{,_cc}
, see #181.CC @hmaarrfk @xhochy @wolfv @conda-forge/tensorflow, feel free to add more things in the description. If any of those points end up being a huge hassle, it can of course be pushed off to the build after that.
Anyone is welcome to chime in if I forgot something. 🙃